PopSGD: Decentralized Stochastic Gradient Descent in the Population Model

The ability to scale distributed optimization to large node counts has been one of the main enablers of recent progress in machine learning. To this end, several techniques have been explored, such as asynchronous and decentralized execution--which significantly reduce the impact of communication and synchronization, and the ability for nodes to perform several local model updates before communicating--which reduces the frequency of communication. In this paper, we show that these techniques, which have so far been considered independently, can be jointly leveraged to obtain near-perfect scalability for training neural network models via stochastic gradient descent (SGD). We consider a setting with minimal coordination: we have a large number of nodes on a communication graph, each with a local subset of data, performing independent SGD updates onto their local models. After some number of local updates, each node chooses an interaction partner uniformly at random from its neighbors, and averages its local model with the neighbor's model. Our first contribution is in proving that, even under such a relaxed setting, SGD can still be guaranteed to converge to local minima under standard assumptions. The proof improves existing techniques by jointly handling decentralization, asynchrony, and local updates, and by bounding their impact. On the practical side, we instantiate this algorithm onto a supercomputing environment, and show that it can successfully converge and scale for large-scale image classification models, matching or even slightly improving the accuracy of the baseline parallel variants.